AI Meeting Assistants Compared: Otter, Fireflies, Notta, Krisp, and FunASR

Meeting assistant marketing is all about “hands-free, auto-summarized.” What you actually hit first in deployment are three more basic things: whether Chinese (especially mixed Chinese-English) meetings transcribe accurately, whether a bot joining the call is allowed and accepted, and where the recordings are stored. Summarization is the easiest part to verify — with correct transcription, summaries rarely go far wrong; with wrong transcription, the best summary is an eloquent error.

This guide compares Otter.ai, Fireflies.ai, Notta, Krisp, and the open-source FunASR, focused on meeting recording, transcription, and summaries. Speech synthesis and voice cloning are a different category — see the AI voice generation comparison.

Quick Verdict

Tool	Correct role	Best for	Main tradeoff
Otter.ai	The English transcription benchmark	English-first meetings, interviews, lectures	Weak Chinese; overseas account required
Fireflies.ai	Meeting bot + workflow integrations	Sales and teams pushing notes into CRM/collab tools	Bot-dependent; privacy optics need managing
Notta	Chinese/Japanese/English multilingual transcription	Chinese and bilingual meetings, interviews, file transcription	Minute and summary quotas vary by plan
Krisp	Device-side noise cancellation + bot-free transcription	Bot-unfriendly meetings, poor call quality	Note-taking depth trails dedicated products
FunASR	Open-source Chinese ASR toolbox	Teams needing private deployment, recordings never leaving the network	Build your own pipeline and summary layer — pure engineering route

In one line: Otter for English meetings, Fireflies to wire notes into CRM and workflows, Notta for Chinese and bilingual meetings, Krisp where bots cannot join, and FunASR self-hosted when data must stay inside your network.

Scope and Method

This article covers meeting recording, transcription, speaker separation, and note generation. Real-time caption hardware, speech synthesis, and voice cloning are out of scope. The built-in transcription in Tencent Meeting and Feishu (Minutes) is a major incumbent option for China-based teams — platform features rather than standalone products — and appears here as a baseline.

The evaluation method is same-recording testing: prepare three of your own real meeting recordings (one Mandarin, one mixed Chinese-English, one multi-speaker crosstalk), run the same files through each product, and manually count word errors and speaker misattributions. Capabilities and plans follow official documentation (access verification attempted 2026-07-24); no prices or minute quotas are pinned.

Chinese Transcription Quality: Test Before Buying

The language gaps among these five are structural, not tunable:

• Notta is designed for the Chinese/Japanese/English market, officially supporting 58 languages and bilingual transcription. For Chinese and mixed-language meetings it is usually the steadiest commercial option — terminology and names still need human review.
• Otter’s models are English-optimized: the most mature English transcription, live highlights, and summaries. Chinese meetings are not its battlefield.
• Fireflies supports multilingual transcription; Chinese works but test accents and domain vocabulary with your own recordings.
• Krisp transcribes mainstream languages, but its core value is noise cancellation — cleaning the audio first measurably improves any downstream transcription.
• FunASR’s Paraformer model family has long been first-tier among open-source Chinese ASR, with VAD, punctuation, and speaker pipelines included; results depend on the model you choose and your compute.

Focus the test on three hard cases: domain terms and product names, mixed-language sentences (“这个 feature 下个 sprint 上线”), and speaker separation during crosstalk.

Capture Method: Bot-Join or Device-Side

This is the fundamental product split, and it directly shapes privacy optics:

Bot joins the meeting (Otter, Fireflies, and Notta all support this): calendar integration auto-joins Zoom/Teams/Google Meet, and participants see “X’s notetaker.” Pros: fully automatic, device-independent. Cons: in external meetings the bot may be refused entry or make the other side uncomfortable; unsuitable for sensitive meetings.

Device-side capture (Krisp’s route; Notta also offers local recording): processing at your computer’s audio layer, no bot visible, works with any meeting software. Pros: low-profile, covers phone calls and in-person scenarios. Cons: captures only your end’s audio path, and compliance responsibility is entirely yours — others see no recording indicator.

File import (all five; FunASR’s primary form): upload recordings after the meeting. The safest fallback, and the only route for historical recordings and voice-recorder material.

Privacy and Compliance: Three Gates Before Recording

Meeting recording implicates participant consent in most jurisdictions; under China’s PIPL, recordings are personal information. Pass three gates before rollout:

1. Consent mechanism: for bot products, confirm join notifications and announcements; for device-side capture, build your own disclosure flow (opening statement, a note in the calendar invite). Put internal meetings into policy; disclose explicitly in every external meeting.
2. Data destination: Otter, Fireflies, Notta, and Krisp all process in the cloud — confirm storage region, retention, deletion, and training use; classify meetings involving customer information or unreleased business data first. Organizations whose recordings must never leave the network have exactly one route: self-hosted FunASR.
3. Access control: default sharing scope of notes, cross-team visibility, and offboarding revocation — a notes tool easily becomes a company-wide meeting-content search engine, which is both the value and the risk.

Export and Workflow Integration

Notes create value in circulation, not in storage. Check four things:

• Export formats: transcript (TXT/DOCX/SRT), summary, and action items exportable separately; SRT export matters for content teams reusing material.
• System push: Fireflies is strongest here — notes, highlights, and action items push into CRMs (Salesforce/HubSpot) and collaboration tools (Slack/Notion); auto-archiving sales calls is its core scenario.
• API and automation: for custom routing (per-project archiving, approval triggers), confirm API capability or relay through a workflow automation platform.
• Search: cross-meeting full-text search with speaker/date filters decides the long-term value of your notes inventory.

The FunASR route builds this layer entirely yourself: transcription output feeds an LLM of your choice for summaries, then your own knowledge base — heavy engineering, full control.

Access and Account Requirements

Otter, Fireflies, and Krisp are overseas SaaS requiring international accounts and payment, with access stability varying by network environment. Notta offers a Chinese interface and a more complete localization path. China-based teams should first check what they already have: the built-in transcription in Tencent Meeting and Feishu requires no new procurement and has the cleanest compliance chain — confirm whether built-ins suffice before evaluating standalone tools. Self-hosted FunASR has no access dependency. For enterprise procurement checks, see the China-accessible AI tools guide.

FAQ

Which transcribes Chinese most accurately?

Among commercial products, Notta is usually the first pick; for private deployment, FunASR’s Paraformer models. But “most accurate” swings with recording quality, accents, and terminology density — test three of your own real recordings on each; half a day settles it.

What accuracy can I expect?

With quiet rooms, standard Mandarin, and no jargon, mainstream products all reach usable levels; noise, dialects, and dense terminology raise error rates sharply. Distrust any accuracy number detached from your recording conditions, and never treat an unreviewed transcript as the basis for meeting decisions.

What if clients refuse the bot?

Switch to the device-side route (Krisp, or local recording plus import), disclose verbally at the meeting open, and obtain consent. Note: technically able to record does not mean compliantly able to record — unconsented recording is both rude and illegal in most scenarios.

Can AI-generated action items be used directly?

As drafts. Common AI summary failures: writing a discussed option as a decision, attributing one person’s suggestion to another. Have the meeting owner verify action items against the transcript before sending — do not skip this step.

Are Tencent Meeting / Feishu built-in notes enough?

For most internal meetings of China-based teams, yes — with the shortest compliance chain. Standalone tools add value in three places: cross-platform meetings (mixed Zoom/Teams), CRM-class integrations, and bilingual/file transcription. Without those needs, don’t add a tool.

How big is the FunASR self-hosting cost?

The models are free; the engineering is not: inference compute (GPU or CPU), deployment operations, a summary layer (attach an LLM), frontend, and permissions all need building. Right for organizations with engineering teams and hard data-compliance constraints. Estimate as recording hours × processing cost + operations labor; low-volume teams are cheaper on commercial products.

Official Sources and Verification

• Otter.ai: otter.ai, access verification attempted 2026-07-24.
• Fireflies.ai: fireflies.ai and its integration directory, access verification attempted 2026-07-24.
• Notta: notta.ai, access verification attempted 2026-07-24.
• Krisp: krisp.ai, access verification attempted 2026-07-24.
• FunASR: GitHub repository and model cards, access verification attempted 2026-07-24.

Language support, minute quotas, and data terms change frequently; this article pins no prices or allowances. Procurement and recording-compliance decisions follow the official terms of the day.

Bottom Line

The selection order for meeting assistants: test Chinese transcription with your own recordings first (the gap is structural), confirm the capture method fits your meeting scenarios (bot-join versus device-side), pass the three privacy gates (consent, data destination, access control), then check that exports plug into your existing workflow. Otter for English, Fireflies for integrations, Notta for Chinese and bilingual, Krisp for bot-free capture, FunASR for data that stays inside — and China-based teams should first confirm whether Tencent Meeting and Feishu built-ins already suffice. Transcription is the foundation; the summary is a thin layer on top — spend your testing budget on the foundation.